Transformers 5.18 ships Nemotron 3 diarization

By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology

♥

The release also clears deprecated Mamba and mask, cache, and rotary paths, while TRL drops vLLM 0.20.0 support.

218 commits 175 PRs merged 5 releases 7 briefings covered

all huggingface reviews →

← previous week huggingface wire →

Release 5.18.0 huggingface/transformers

by vasqu

Nemotron 3 Diarization is an open-weight streaming speaker diarization model, released alongside the usual model additions in the 5.18.0 tag. It is a breaking release, so treat the version bump as a migration, not a drop-in.

Seed before creating the model in experimental trainers huggingface/trl

by albertvillanova

The experimental trainers build the PEFT model in __init__ before Trainer.__init__ calls set_seed, so LoRA adapters were getting whatever RNG state happened to be current. The trainer now seeds before creating the model, matching what the main trainers already do.

Drop vLLM 0.20.0 support (#7435) huggingface/trl

by Quentin Gallouédec

vLLM 0.20.0 is no longer supported, which lands as a hard floor for anyone running TRL's async or vLLM-backed trainers. Pin your vLLM version deliberately rather than inheriting whatever the environment resolves.

Add selective activation checkpointing (SAC) to SFT huggingface/trl

by kashif

With gradient checkpointing on, attention output is now saved during forward instead of recomputed in backward, recovering most of the checkpointing slowdown at long context for one extra hidden-state-sized tensor per layer.

Deprecate opencode_env and pi_env in favour of harbor_env huggingface/OpenEnv

by sergiopaniego

OpenCode and Pi are validated Harbor harnesses with the same token-level capture for training, so the per-harness envs duplicate an existing path. Both are removed in OpenEnv 0.8.0, which is your migration window if you reference opencode_env or pi_env.

Deprecate the use_mamba_kernels config flag that no longer has any effect (#49155) huggingface/transformers

by Albert Villanova del Moral

The flag stays accepted in configs but no longer has any effect, so existing configs will not break, they will just keep carrying a setting that does nothing.

fix(setup): give xet-core duration defaults a unit so they apply huggingface/hf-mount

by XciD

xet-core parses durations with humantime, which rejects a bare number and silently falls back to its own default with a WARN hidden by the default hf_mount=info filter. Two defaults set in init_tracing never applied as a result, so hf-mount users should recheck their duration settings.

[MFU] Use device peak flops instead of hardcoded H100's huggingface/trl

by dwarez

Async trainers computed MFU with the hardcoded H100 peak flops default; they now map peak flops by device, so the number reflects the hardware you actually ran on. Alongside: kernels removed the cxx11 ABI tag from Linux builds while keeping client-side parsing of tagged variants, funes now pulls unpushed local memory turns into recall and brings installed integrations forward on update, and Rep...

$ ls huggingface/week/ # the briefings behind this review

Mon Sep 28 TRL seeds before adapter init, docs links repaired Tue Sep 29 Repo2RLEnv learns mocha's output format, transformers cleans up MPS and Mamba leftovers Wed Sep 30 hf-mount duration defaults silently ignored, funes update now bumps integrations Thu Oct 1 Transformers 5.18 ships Nemotron 3 diarization, TRL drops vLLM 0.20.0 Fri Oct 2 transformers clears out deprecated mask, cache, and rotary paths Sat Oct 3 QLoRA keeps DoRA in float32, flash-attention kwargs stop leaking into vision encoders Sun Oct 4 AnyLanguageModel catches up to Foundation Models 26 and 27

all huggingface reviews →

← previous week huggingface wire →